Professional Documents
Culture Documents
net/publication/357876127
CITATIONS READS
0 1,085
4 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Ting Tin Tin on 17 January 2022.
Abstract. This research is a study on the young generation views and acceptance of Artificial Intelligence
(AI) art based on the painting and literature created by the latest AI technologies to understand AI
advancement and capabilities in the domain of art. An online questionnaire is sent to a university’s
undergraduates in a local university with 202 responds. Data is analyzed by PSPP utilizing Cronbach’s Alpha
to test the reliability of the questionnaire items. Frequencies, Mean, Chi-Square, and T-Test are used in
analyzing the hypotheses. The results show that 54% of respondents did not correctly identify emotions in AI
artworks and therefore could not appreciate the emotion expressed in AI artworks. There was no correlation
between degree of exposure to AI technology and acceptance of AI artworks although the positive
acceptance rate is high - 74%. An average of 55.3% of the respondents are able to correctly differentiate
between AI artwork and human artwork. Lastly, there is no correlation between attributed artist identity and
judgement on AI images but there is a correlation for AI poem.
INTRODUCTION
The term Artificial Intelligence (AI) was originated in the 1950’s in modelling human cognition. Nowadays, the
term has evolved to refer to application that rely on deep neural networks [7]. AI art refers to artwork made by
collaboration between AI algorithms and human artists. There are several common perceptions regarding AI art.
Firstly, there is a misconception that AI will never excel in creative arts such as painting and literature and thus will
never achieve creativity. This also means that humans have the capability to differentiate between artworks
produced by AI versus humans. Therefore, the public should be educated on the roles of AI in artworks and AI
creativity though AI might not be able to fully replace human’s creativity. On the other hand, humans might not be
able to appreciate AI artworks emotionally as they treated it as “artificial”. In the previous studies, there are only
recordings of human judgements on AI artworks to evaluate how good an AI artworks algorithm is. However, not
many focused on how an attributed artist identity might affect the judgement on AI art. Hence, these are the areas
that this study wants to investigate. This paper is constructed in four main sections: Literature Review, Research
Methodology, Research Results and Discussion, and Conclusion.
LITERATURE REVIEW
What are the roles of AI in art? According to previous studies, AI can be utilized as Imitator, Collaborator, and
Creator. AI as an Imitator implements the concept of style transfer, which involves the imitation and mixing of
existing human artworks [14, 15, 16, 19]. In visual arts style transfer, image-based artistic rendering (IB-AR)
techniques are either with or without Convolutional Neural Networks (CNNs) [1, 29, 30]. Techniques without the
use of CNNs are typically limited in flexibility, style diversity and effective image structure extractions. In contrary,
Neural Style Transfer (NST) utilizes CNNs to extract content information from an image and style information from
an artwork. Such content and style information are then combined again to produce the image in the style of the
artwork [21]. NST has overcome limitations of non-CNN based techniques because it is not restricted to a limited
range of styles.
While AI as an Imitator strives to combine an existing image with an existing art style, the goal of AI as a
Collaborator is to create novel artworks itself, though with heavy human involvement. In 2018, a painting as the
product of human-AI collaboration was auctioned for $432,500 (approximately RM1,716,022) at a British auction
house [2]. Technically, the painting was created using Generative Adversarial Networks (GANs) [1]. GANs consist
of two algorithms in competition with one another: a generator to produce novel content after being fed with training
data, and a discriminator to differentiate the original artworks from the generator’s output [23]. The generator
tweaks its output until the discriminator can no longer find any difference. In the case of this painting, human
intervention comes in at the selection stage, where image outputs were selected based on the human curators’ tastes.
One advantage of AI art algorithms is that they are constantly learning and improving based on past experiences and
training data. Hence, the human-AI collaboration can be a means to maximize both parties’ creative capacities in
arts [34].
AI as a Creator means that other than human effort invested during the process of its learning, no human
intervention should occur during its art creation stage, which should eventually produce truly novel artworks [17].
Elgammal’s research has dismissed the ability of GANs to produce creative outputs. This is because GAN
generators will, under the supervision of the discriminator, ultimately tweak its outputs to an extent that they look
like existing artworks. The generator is not pushed to explore the creative conceptual space. Since art creation is
tightly knitted with arousal potential of the human emotion [20, 33], GANs, failing at creating artworks that would
arouse excitement, cannot produce creative artworks. Therefore, Elgammal proposed a new art-generating system
named Creative Adversarial Network (CAN) [17]. CAN is an adaptation of GANs where it also uses the generator
and discriminator networks. The difference lies in the training, where the discriminator instead of generator is
trained to recognize the styles of artworks. The generator is not fed with any training data but is set to produce
random outputs. Then it will receive two feedbacks from the discriminator: whether the output is an art, and how
easily the discriminator sorts it into an existing art style. The generator will be rewarded or penalized based on the
discriminator’s feedback.
The leading definition that is undisputed among researchers for creativity is the ability to generate novel (new)
and valuable (useful) ideas [8, 9, 22, 35, 38]. Novelty is categorized into psychological (P-creativity) and historical
novelty (H-creativity). P-creativity refers to the generation of an idea that is merely new to the entity that has
generated it, whereas H-creativity produces one that is entirely unseen or unheard of in human history. The painting
machine AARON that produced vividly original works pursued H-creativity. Its creator Cohen even elevated
AARON to the height of a “world-class” colorist [40]. However, computer models most often aim for P-creativity
[6]. According to Boden, there are three types of creativity: combinational, exploratory and transformational
creativity [4].
Combinational creativity generates new combinations from familiar ideas by identifying indirect associations
between two concepts from random sources [25]. For example, Poem.exe generates and posts unique Japanese
Haiku poems on Twitter through random combinations of data input. However, AI algorithms face challenges in
combinational creativity. There are two aspects to creativity: novelty and value. While Poem.exe can produce new
poems every day, many times it generates nonsensical texts. A valuable result should be both interesting and
sensible [31, 39]. How can AI art algorithms evaluate the value of their outputs? A possible solution is to introduce
creative autonomy for AI machines, so that they evaluate creations independently and change generative standards
without explicit direction [28].
Exploratory creativity is where the style and rules are constrained within the conceptual space, but continual
learning is needed to seek the space’s potential and limits. However, how should AI machines learn? Turing
suggests that machines can be developed to simulate a child’s mind rather than an adult’s [11]. AARON is an
epitome of child-learning, seeing that its painting style from 1988 through 2013 had drastically changed. Cohen
admits that AARON was like a young child learning how to draw [10].
Transformational creativity includes the alteration of the dimensions that defines the conceptual space,
particularly involving the breaking of rules. How can AI art algorithms break rules? Genetic Algorithms (GAs) can
be incorporated into the program, which allows random changes that are similar to mutations in biological evolution
[5]. In order to avoid radically different changes that causes a discontinuation of style, sculptor William Latham
cautiously and sparingly employed GAs in his AI program, eliminating large point mutations altogether. As a result,
the series of produced images was artworks beyond himself but remains similar in his artistic style [41].
Can AI excel in artwork production? Back in 1985, AI still lacked of human expression in one of the arts, music-
making. However, after decades of evolution, AI is now widely used in audio mastering, even helping students to
learn music [3, 32, 36]. Despite huge advancements in AI art, it seems like humans are unable to appreciate or
apprehend AI artworks. For example, in the emotion connection to AI artwork researches, participants found that
ePainter’s paintings is too emotionally depressing [37]. Generative Pre-trained Transformer (GPT) generates
writings that still lack of semantic coherence and is difficult to understand [18, 11]. Based on these researches, it is
assumed that human still cannot identify emotions in AI artworks. In a study of the “mere exposure” effect, it was
found that repeated mere exposure to something will enhance an individual’s attitude towards it [12, 24]. Based on
this effect, it is assumed that exposure to AI technology will affect one’s acceptance of AI artworks. Therefore, this
study will focus in undergraduates with higher possibility in AI technologies knowledge exposure.
Finally, bias might affect human judgement on AI art. Reception of AI paintings was positive when human
subjects were unaware of the painter’s identity [24, 26] When people were told certain paintings were created by AI
(attributed artist identity = AI), they rated the paintings significantly lower than other people who thought the artist
was human (attributed artist identity = human) in terms of composition, degree of expression and aesthetic value
[27]. Hence, we want to replicate the experiment and find out how attributed artist identity might affect the
judgement on AI art.
RESEARCH METHODOLOGY
Online questionnaire is selected to collect data due to the COVID-19 pandemic. The questionnaire consists of
five sections. Section 1 is about demographic background of respondents which includes respondent’s age, gender
and faculty. Section 2 collect general impression of respondents on AI artworks in terms of emotions. Four AI
paintings that look the same but with different color palettes were used in the questions. Respondents selected one
from four phrases that best described the painting’s emotions (Figure 1). These four paintings are adapted from
Salevati and DiPaola [13, 37].
Section 3 is a Turing test which is used to test whether respondents could pick the correct artwork produced by
AI [11]. Three pairs of paintings (Figure 2) and one pair of poems were used. In each pair, one was produced by AI
and another was by human. Respondents are required to identify artwork produced by AI.
1. AI Artwork 2. Human Artwork
Section 4 is about personal opinion which elicit the degree of exposure to AI technology and the respondents’
acceptance of AI artworks. The questions to determine the acceptance variable were taken from the research done by
Hong and Curran [26, 27]. These four sections (one – four) are non-experimental design. Section 5 was an
experimental design to find out the relationship between attributed artist identity and respondents’ judgement on AI
artworks. This design was modified from Hong and Curran’s research [26, 27]. Section (5) has five questions that
involved three paintings and two writings produced by AI. In questionnaire set A, respondents were told that these
artworks were produced by AI. In questionnaire set B, respondents were told that these artworks were produced by
human. Respondents were required to rate the artworks on a five-point Likert scale on several evaluation criteria.
Evaluation criteria were inspired by Hong and Curran’s research [26, 27] and Elkins and Chan [18].
FIGURE 3. Examples of AI artworks used in Section 5 modified from Hong and Curran’s research [26, 27].
Based on the literature review, this study’s hypotheses are constructed as follows:
H1: Undergraduates are not able to identify emotions in AI artworks.
H2: There is a correlation between degree of exposure to AI technology and acceptance of AI artworks.
H3: Undergraduates are able to correctly identify human artwork versus AI artworks.
H4: There is a correlation between attributed artist identity and judgement on AI artworks.
Demographics
Table 1 shows a total of 202 respondents demographics: age, gender, and faculty (Questionnaire Section 1).
From the data, most of respondents are male from FOCS as there are majority of male in the faculty.
TABLE 1. Respondents’ demographics data.
No. of Respondents Percentage (%)
Age <18 1 0.5
18-20 85 42.1
21-25 115 56.9
>25 1 0.5
Gender Male 106 52.5
Female 96 47.5
Faculty FOCS 111 55
FAFB 51 25.2
FOET 2 1
FSSH 14 6.9
FCCI 8 4
FOBE 6 3
FOAS 10 5
FOCS: Faculty of Computing and Information Technology
FAFB: Faculty of Accountancy, Finance and Business
FOET: Faculty of Engineering and Technology
FSSH: Faculty of Social Science and Humanities
FCCI: Faculty of Communication and Creative Industries
FOBE: Faculty of Built Environment
FOAS: Faculty of Applied Sciences
H2: There is a correlation between degree of exposure to AI technology and acceptance of AI artworks.
Based on Table 3, respondents have moderate exposure to AI technology (Mean = 3.33). Table 4 shows the
acceptance rate of AI artworks and the mean is 74.9%. However, based on Table 5, there is no significant correlation
between degree of exposure to AI technology and acceptance of AI artworks. Therefore, H2 is rejected. This result
is different from the theory proposed by Zajonc’s “Mere Exposure Effect” [12].
TABLE 5. Ch-square test for the relationship between acceptance rate and degree
of exposure.
Pearson Chi-Square Asymptotic Sig. (2-tailed)
Q1 x Rate 1.92 0.751
Q2 x Rate 2.65 0.619
Q3 x Rate 17.14 0.144
Q4 x Rate 9.17 0.689
H3: Undergraduates are able to correctly identify human artwork versus AI artworks.
Section 3 is a Turing test based on 8 images. Respondents are required to compare two images created by AI
artwork and human artwork [11]. Based on Table 6, an average of 55.3% of the respondents are able to correctly
differentiate between AI artwork and human artwork. Therefore, H3 is accepted.
TABLE 6. Turing test – differentiate between AI artwork and human artwork.
Total Number of Respondents Total Percentage of Respondents
identify correctly identify correctly
Q1. Image 1 & 2 75 37%
Q2. Image 3 & 4 126 62%
Q3. Image 5 & 6 134 66%
Q4. Image 7 & 8 83 41%
MEAN 55.3%
H4: There is a correlation between attributed artist identity and judgement on AI artworks.
In this experimental test, respondents are informed that the images in a subsection are created by AI.
Respondents have to judge the images accordingly. In the next subsection, AI artworks are shown to the respondents
but respondents are informed that these images are produced by human artists. For each attributed identity group, the
human artist identity group consistently gain slightly higher ratings compared to the AI artist identity group (Table
8). Based on the independent Sample T-Test, there is a relationship between attributed artwork and judgement on AI
artwork for Poem only (Table 7). Since not all evaluation criteria have significant differences in means, therefore H4
is rejected.
CONCLUSION
This study has combined several experiments from previous studies regarding human attitudes and perceptions
towards AI artworks. The findings are interesting in which most of the result did not comply with previous studies.
For example, there is no relationship between exposure of AI and acceptance of AI artworks. There is also no
relationship between attributed artwork and judgement on AI artwork. This has proven that researchers can focus in
improving AI algorithm in order to produce more realistic artworks as the acceptance and judgement of AI artwork is
not based on one’s AI knowledge or exposure to AI as arts is a field of feelings and it depends on different
individuals’ experience and personal preference.
REFERENCES
1. Anantrasirichai, N. and Bull, D. (2021) Artificial Intelligence in the Creative Industries: A Review. Artificial
Intelligence Review (2021). https://doi.org/10.1007/s10462-021-10039-7.
2. Arbiza Goenaga, M. (2020) A Critique of Contemporary Artificial Intelligence Art: Who is Edmond de Belamy?
AusArt 8(1), pp. 51–66. https://doi.org/10.1387/ausart.21490.
3. Birtchnell, T. and Elliott, A. (2018) Automating the Black Art: Creative Places for Artificial Intelligence in
Audio Mastering. Geoforum 96(August), pp. 77–86. https://doi.org/10.1016/j.geoforum.2018.08.005.
4. Boden, M. A. (2004) The Creative Mind: Myths and Mechanisms: Second Edition. Routledge.
https://doi.org/doi: 10.4324/9780203508527.
5. Boden, M. A. (2006) Mind as Machine: A History of Cognitive Science. Minds and Machines 18(1), pp. 121–
125. https://doi.org/10.1007/s11023-008-9091-9.
6. Boden, M. A. (2009) Computer Models of Creativity. AI Magazine 30(3), pp. 23–34.
https://doi.org/10.1609/aimag.v30i3.2254.
7. Cardon, D., Cointet, J. P. and Mazières, A. (2018) Neurons Spike Back: The Invention of Inductive Machines
and the Artificial Intelligence Controversy, Reseaux. doi: 10.3917/res.211.0173.
8. Chen, L., Wang, P., Shi, F., Han, J. and Childs, P. R. N. (2018) A Computational Approach for Combinational
Creativity in Design. In Proceedings of the DESIGN 2018, pp. 1815-1824.
https://doi.org/10.21278/idc.2018.0375.
9. Childs, P. R. N. (2019) Mechanical Design Engineering Handbook. Elsevier. https://doi.org/10.1016/C2016-0-
05252-X.
10. Cohen, H. (1982) How to Make a Drawing. In talk given to the Science Colloquium, National Bureau of
Standards, Washington DC (vol. 17).
11. Turing, A. M. (1950) Computing Machinery and Intelligence. Mind.
12. Zajonc, R. B. (1968) Attitudinal Effects of Mere Exposure. Journal of Personality and Social Psychology, 9(2
Part 2), pp. 1–27. https://doi.org/10.1037/h0025848.
13. Dipaola, S., Gabora, L. and McCaig, G. (2018) Informing Artificial Intelligence Generative Techniques Using
Cognitive Theories of Human Creativity. Procedia Computer Science, 145, pp. 158–168. https://doi.org/
10.1016/j.procs.2018.11.024.
14. Drori, I., Cohen-Or, D. and Yeshurun, H. (2003) Example-based Style Synthesis. In Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern Recognition. https://doi.org/
10.1109/cvpr.2003.1211464.
15. Efros, A. A. and Freeman, W. T. (2001) Image Quilting for Texture Synthesis and Transfer. In Proceedings of
the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 2001.
https://doi.org/ 10.1145/383259.383296.
16. Elad, M. and Milanfar, P. (2017) Style Transfer Via Texture Synthesis. IEEE Transactions on Image Processing.
https://doi.org/10.1109/TIP.2017.2678168.
17. Elgammal, A., Liu, B., Elhoseiny, M. and Mazzone, M. (2017) CAN: Creative Adversarial Networks Generating
"Art" by Learning About Styles and Deviating from Style Norms. In Proceedings of the 8th International
Conference on Computational Creativity.
18. Elkins, K. and Chun, J. (2020) Can GPT-3 Pass a Writer’s Turing Test? Journal of Cultural Analytics.
https://doi.org/10.22148/001c.17212.
19. Frigo, O., Sabater, N., Delon, J. and Hellier, P. (2016) Split and Match: Example-based Adaptive Patch
Sampling for Unsupervised Style Transfer. In Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, pp. 553-561. https://doi.org/10.1109/CVPR.2016.66.
20. Gardner, H. and Berlyne, D. E. (1974) Aesthetics and Psychobiology. Curriculum Theory Network.
https://doi.org/10.2307/1179240.
21. Gatys, L., Ecker, A. and Bethge, M. (2016) A Neural Algorithm of Artistic Style. Journal of Vision.
https://doi.org/10.1167/16.12.326.
22. Gaut, B. (2010) The Philosophy of Creativity. Philosophy Compass. https://doi.org/10.1111/j.1747-
9991.2010.00351.x.
23. Ian, G., Pouget-Abadie, J., Mirza, M., Xu, B. and Warde-Farley, D. (2014) Generative Adversarial Nets. In
Advances in Neural Information Processing Systems, pp. 2672-2680. https://doi.org/10.3156/jsoft.29.5_177_2.
24. Inoue, K., Yagi Y. and Sato N (2017) The Mere Exposure Effect for Visual Image. Mem Cogn (2018) 46:181–
190, https://doi.org/10.3758/s13421-017-0756-6.
25. Han, J., Shi, F. and Childs, P. R. N. (2016) The Combinator: A Computer-based Tool for Idea Generation. In
Proceedings of International Design Conference.
26. Hong, J. (2018) Bias in Perception of Art Produced by Artificial Intelligence. https://doi.org/10.1007/978-3-319-
91244-8.
27. Hong, J. W. and Curran, N. M. (2019) Artificial Intelligence, Artists, and Art: Attitudes Toward Artwork
Produced by Humans vs. Artificial Intelligence. ACM Transactions on Multimedia Computing,
Communications and Applications, 15(2s). https://doi.org/10.1145/3326337.
28. Jennings, K. E. (2010) Developing Creativity: Artificial Barriers in Artificial Intelligence. Minds and Machines,
20(4), pp. 489–501. https://doi.org/10.1007/s11023-010-9206-y.
29. Jing, Y., Yang, Y., Feng, Z., Ye, J., Yu, Y. and Song, M. (2018) Neural Style Transfer: A Review. IEEE
Transactions on Visualization and Computer Graphics, 26(11), pp. 3365–3385. https://doi.org/10.1109/
TVCG.2019.2921336.
30. Kolliopoulos, A. (2005) Image Segmentation for Stylized Non-Photorealistic Rendering and Animation.
University of Toronto, pp. 0395-0395.
31. Kurt, D. E. (2018) Artistic Creativity in Artificial Intelligence. p. 87.
32. López de Mántaras, R. (2016) Artificial Intelligence and the Arts: Toward Computational Creativity. Open
Mind.
33. Martindale, C. (1990) The Clockwork Muse: The Predictability Of Artistic Change. New York: Basic Books.
34. Mazzone, M. and Elgammal, A. (2019) Art, Creativity, and the Potential of Artificial Intelligence. Arts, 8(1), p.
26. https://doi.org/10.3390/arts8010026.
35. Pereira, F. C. (2007) Creativity and Artificial Intelligence: A Conceptual Blending Approach. Walter De Gruyter
(Vol. 4). https://doi.org/10.1515/9783110198560.
36. Roads, C. (1985) Research in Music and Artificial Intelligence. ACM Computing Surveys (CSUR), 17(2), pp.
163–190. https://doi.org/10.1145/4468.4469.
37. Salevati, S. and DiPaola, S. (2015) A Creative Artificial Intelligence System to Investigate User Experience,
Affect, Emotion and Creativit. pp. 140–147. https://doi.org/10.14236/ewic/eva2015.13.
38. Sarkar, P. and Chakrabarti, A. (2011) Assessing Design Creativity: Refinements to the Novelty assessment
Method, Design Studies, 32(4), pp. 348–383. https://doi.org/10.1016/j.destud.2011.01.002.
39. Sawyer, R. K. (2006) Explaining Creativity: The Science of Human Innovation. New York: Oxford University
Press.
40. Sundararajan, L. (2014) Mind, Machine, and Creativity: An Artist’s Perspective. Journal of Creative Behavior,
48(2), pp. 136–151. https://doi.org/10.1002/jocb.44.
41. Todd, S. and Latham, W. (1992) Evolutionary Art and Computers. London: Academic Press.